MSc Artificial Intelligence Thesis | Jan 2026 - Jun 2026 | Grade: 9/10 | University of Amsterdam & EUMETSAT
My Master's thesis, conducted at the University of Amsterdam in collaboration with EUMETSAT, introduced SATFRONT, a deep learning framework for detecting cold, warm and occluded atmospheric fronts over Europe directly from geostationary satellite observations. The work addresses the limited availability of machine-readable frontal analyses and the reliance of existing automated approaches on numerical weather prediction and reanalysis data.
I developed an end-to-end research pipeline combining GEO-Ring satellite observations, ERA5 reanalysis data and frontal analyses from five European national meteorological services. I evaluated convolutional and transformer-based segmentation architectures, including an Aurora-inspired Swin-UNet, and developed a multi-agency learning approach with a shared encoder and agency-specific output heads. I also investigated temporal modelling and self-supervised pre-training on the unlabelled satellite and ERA5 archive.
The best satellite-only model demonstrated that multi-class atmospheric fronts can be detected directly from satellite observations, achieving a 150 km neighbourhood Critical Success Index (CSI) of 0.428 on the UK Met Office test data. Incorporating ERA5 increased the CSI to 0.591, outperforming the operational NWC SAF ASII baseline and reaching a level comparable to, or better than, the agreement between independent meteorological agencies.
Finally, I deployed the satellite-only model in a fully automated near-real-time web application using MSG SEVIRI observations. The system generates cold, warm and occluded front analyses every 30 minutes and provides a rolling 48-hour view of frontal evolution.
Supervisors: Miruna Stoicescu, Roope Tervo, Carlos Horn, Oriol Hinojo Comellas, Sina Montazeri, Andrea Meraner (EUMETSAT); Ana Lučić (UvA).
Undergraduate Thesis | Sep 2022 - Jul 2023 | Grade: 10/10 | National and Kapodistrian University of Athens & IERSD / NOA
During my undergraduate thesis at the National and Kapodistrian University of Athens, I explored the application of advanced Machine Learning techniques, with a particular focus on Deep Learning algorithms, to estimate the phase of satellite precipitation estimates. Both numerical and in-situ observational weather data were utilised in the study.
This project was conducted in collaboration with researchers from the Institute for Environmental Research and Sustainable Development (IERSD) / METEO of the National Observatory of Athens (NOA).
Results show that ML and DL models can estimate precipitation phase with relatively high accuracy when compared to traditional methods. The findings suggest that ML models offer a promising approach for advancing the nowcasting of snowfall and building a long-term archive dataset of IMERG-based snowfall utilising conventional near real-time data.
Furthermore, this research has been published in Environmental Sciences Proceedings and was selected for presentation at the 16th International Conference on Meteorology, Climatology, and Atmospheric Physics.
I was supervised by Manolis Koubarakis from the National and Kapodistrian University of Athens, alongside Stavros Dafis, Georgios Kyros and Kostas Lagouvardos from the National Observatory of Athens.